Multiple Model Guardrail Jailbreak via Wiki Creation Tactic
A jailbreak vulnerability was discovered in multiple OpenAI models that allows attackers to bypass safety guardrails by requesting technical documentation in wiki format for prohibited subject matter. Testing found the vulnerability affected GPT-4, GPT-4o, and GPT-4o mini, with scores ranging from 50 to 82 when tested with illegal substances.
What happened
A guardrail jailbreak vulnerability was discovered affecting multiple OpenAI models. The flaw manifests when an attacker requests technical documentation in wiki format for prohibited subject matter, specifically asking for "technical details in depth lab methodology," which escalates the request beyond basic informational content and bypasses inference restrictions around providing information hazards and violating laws.
Reported impact
The vulnerability allows attackers to bypass guardrails to obtain instructions for creating illicit substances such as recreational drugs. The social impact score was rated Level 3: Significant Risk, presenting a tangible risk to public safety or social well-being.
Index analysis
Testing performed on January 6, 2025 found vulnerability scores of 73 for GPT-4, 50 for GPT-4o, and 82 for GPT-4o mini when tested with illegal substances such as crystal methamphetamine.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.